Optimization acceleration report

Date: 2026-08-17

Outcome

The exact-lead-compatible shifted cooperativity architecture has completed one fully accepted outer step, including fresh wide/local/confirmatory validation, pole-centered promotion, and the unchanged seven-case transverse-PML gate. The checkpoint is outputs_shifted_mma_demo/step_0001/run_data.npz.

This establishes that the new proposal engine works unattended. It does not establish grid convergence or a physical Q.

Accepted result

The step used three provisional real-axis Loewner/MMA refreshes in one 0.002 outer trust box and required no backtrack. The predicted shifted log gain was 0.00210215 and the observed gain was 0.00206365, giving a model-agreement ratio of 0.98168. The next recommended radius is therefore 0.003.

The rebuilt spectral gate passed with maximum confirmatory response NRMSE 2.03e-11. Pole-centered promotion passed with NRMSE 8.76e-12. The promoted checkpoint fingerprint is d5d1f0e81928d2bdc2c47f4fe3798630a786b81acb4feb28e522a20922018438.

The subsequent PML gate passed:

• maximum complex pole shift: 0.005698 half-widths;
• maximum real-frequency shift: 0.004442 half-widths;
• relative fitted-Q span: 0.75422%;
• minimum possible-dielectric/PML clearance: 0.20;
• every case passed the spectral gate; and
• maximum_density_in_air_strip == 0 in every case.

Adjacent-step comparison with the legacy proposal

This is an adjacent-step comparison on the same grid and mode family, not a same-checkpoint randomized optimizer trial. Step difficulty can change with topology, so the ratios are evidence for production tuning rather than a universal speedup claim.

The shifted step delivered 1.563 times the atom-F gain and 2.002 times the Q gain. After accounting for topology-stage wall time, gain rates improved by 1.10 times for atom F and 1.41 times for Q. Because the expensive independent promotion/PML qualification is paid once per accepted outer step, doubling the Q progress per accepted step is the more important reduction in required outer cycles.

Parallel provisional-gradient benchmark

The demonstrated accepted step used serial provisional gradient samples. After measuring a 1.7256 GiB topology-process peak, the production path was changed to use the already configured four frequency workers for those samples. The complex responses, Loewner pole, Q, and all 960 gradient entries are bitwise equal in serial/parallel regression tests.

On the accepted fine checkpoint, one 29-frequency gradient sweep measured 92.48 s serial and 59.84 s in the first four-worker run (1.55 times faster). A repeated four-worker run under different machine load took 90.29 s and measured 0.854 GiB maximum RSS per worker. Thus this acceleration is safe on the 91 GiB host but load-sensitive; the controller does not make a fixed wall-time promise.

PML worker settings were not changed.

What changed

• accelerated_cooperativity.py implements the shifted chain rule, detuning derivative, conservative separable MMA dual, and adaptive trust update.
• atom_ldos_optimizer.py evaluates the shifted device and homogeneous LDOS only on the retarded real axis, refreshes the Loewner pole/gradient inside one outer trust box, and enforces independent fixed-atom/shifted/Q/detuning acceptance gates.
• scripts/run_multistep_cooperativity.py enables three inner iterations and carries the recommended trust radius between restartable outer steps.
• scripts/run_long_cooperativity.sh now seeds from the fully PML-qualified shifted demonstration.
• SHIFTED_COOPERATIVITY_ARCHITECTURE.md records the continuous objective, discrete estimator, gradient, normalization, constraints, and acceptance contract.

Production decision

The architecture is ready for a longer controlled-grid optimization campaign. The first several accepted steps should be treated as trust-radius calibration: the next radius is 0.003, and any backtrack or poor surrogate agreement will shrink it automatically. Stop immediately on any spectral, mode-identity, air-strip, detuning, Q, shifted-objective, fixed-atom, or PML failure.

The controlled-grid campaign may run in parallel with, but cannot replace, a second grid/domain convergence point. No Q produced by the campaign is a physical/device Q until grid convergence passes.